Combine photosynthetic characteristics and leaf hyperspectral reflectance for early detection of water stress

Linbao Li, Guiyun Huang, Jinhua Wu, Yunchao Yu, Guangxin Zhang, Yang Su, Xiongying Wang, Huiyuan Chen, Yeqing Wang, Di WuFrontiers in Plant Science 2025 · DOI

This study, recently published in Frontiers in Plant Science, explores a high-tech "early warning system" for trees facing drought. By combining hyperspectral leaf reflectance with machine learning, researchers were able to detect water stress in Rhamnus leptophylla—a critical species for ecological protection in China’s Three Gorges area—well before any visible signs of wilting appeared.

TL;DR

Researchers combined leaf hyperspectral reflectance with machine learning (Random Forest) to non-destructively estimate photosynthetic health and detect water stress in trees. The study identifies specific spectral signatures that act as "pre-visual" indicators of drought, offering a scalable path for precision forest management and remote sensing.

What Problem Does This Paper Address?

Drought is a primary limiter of plant growth, but by the time a tree looks "thirsty" (yellowing or wilting leaves), significant physiological damage has often already occurred. Traditional monitoring—like measuring soil moisture or manual leaf sampling—is either too localized or too slow for large-scale forest management.

The authors sought to bridge this gap by using hyperspectral reflectance: measuring how leaves reflect light across hundreds of narrow bands in the electromagnetic spectrum. The goal was to see if these spectral "fingerprints" could accurately predict internal photosynthetic traits, specifically Leaf Chlorophyll Content (LCC) and Chlorophyll Fluorescence (ChlF), which shift at the onset of water stress.

Methods

The team conducted a controlled greenhouse experiment on three-year-old Rhamnus leptophylla trees, subjecting them to a 13-day drought followed by a 3-day re-watering period.

  • Data Collection: They measured three distinct types of data:
  • LCC: Using a handheld SPAD meter.
  • ChlF: Using a pulse-amplitude-modulated (PAM) fluorometer to track parameters like ETR (electron transport rate) and NPQ (non-photochemical quenching).
  • Hyperspectral Reflectance: Using a spectrometer covering the 325–1075 nm range.
  • Feature Engineering: They calculated 61 different hyperspectral vegetation indices (VIs) and extracted "three-edge" parameters (red, blue, and yellow edges of the spectrum).
  • Modeling: They compared standard Multivariate Linear Regression (MLR) against three machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN).

Datasets

The study utilized a primary experimental dataset generated by the authors:

  • Sample Size: 1,134 total samples derived from nine trees.
  • Stratification: Measurements were taken across three canopy layers (upper, middle, and lower) to account for vertical variations in leaf maturity and light exposure.
  • Data Split: 80% for training and 20% for testing, validated with 10-fold cross-validation.

Key Insights and Ideas

  • Fluorescence is the "Canary in the Coal Mine": Chlorophyll Fluorescence (ChlF) parameters were more sensitive to early drought than total chlorophyll content. While LCC stayed relatively stable in the first few days, fluorescence traits like NPQ and ETR shifted almost immediately.
  • The "Sweet Spot" Bands: Reflectance in the 540–560 nm (green) and 750–1100 nm (near-infrared) ranges were the most effective at distinguishing water stress levels.
  • Canopy Matters: The middle canopy layer provided the most accurate LCC estimations ($R^2 = 0.902$). This suggests that sampling strategy—where you point the sensor—is just as important as the sensor itself.
  • RF Wins on Complexity: Random Forest significantly outperformed the other models. Its ability to handle the high dimensionality and multi-collinearity of hyperspectral data allowed it to capture non-linear relationships that linear regression missed entirely.

Key Results

  • LCC Prediction: The Random Forest model achieved an impressive $R^2$ of 0.895 for all chlorophyll samples.
  • ChlF Prediction: RF also dominated here, achieving high accuracy for key traits:
  • $R^2 = 0.854$ for NPQ (stress dissipation).
  • $R^2 = 0.879$ for Y(II) (quantum yield).
  • $R^2 = 0.878$ for ETRmax (maximum electron transport).
  • Early Detection: The Simple Ratio index SR(752/690) was identified as a particularly sensitive tracker, showing significant changes by day 4 of drought—long before visual symptoms emerged.

Limitations and Future Work

  • Species Specificity: The models were trained specifically on Rhamnus leptophylla. While the methodology is robust, the specific spectral weights might vary for other tree species.
  • Controlled vs. Wild: The study was performed in a greenhouse. Real-world variables like varying sun angles, atmospheric interference, and mixed-species canopies present additional challenges for satellite or UAV-based scaling.
  • Next Steps: The authors suggest that the next logical step is integrating these RF models into drone-based hyperspectral platforms to monitor entire forest stands in real-time.

Summary updated

Abstract

Advanced techniques capable of early and non-destructive detection of the impacts of water stress on trees and estimation of the underlying photosynthetic capacities on larger scale are necessary to meet the challenges of limiting plant growth and ecological protection caused by drought. We tested influence of continuous water stress on photosynthetic traits including Leaf Chlorophyll content (LCC) and Chlorophyll Fluorescence (ChlF) and combined hyperspectral reflectance as a high-throughput approach for early and non-destructive assessment of LCC and ChlF traits in <i>Rhamnus leptophylla</i> trees. LCC and ChlF parameters (NPQ, Fv'/Fm', ETR, ETRmax, Fm', qL, qP, Y(II) were measured alongside leaf hyperspectral reflectance from <i>Rhamnus leptophylla</i> suffering from constant drought during water stress. Water stress caused NPQ, Fv'/Fm', ETRmax, Fm', qL, qP, Y(II) and ETR continuous decline throughout the entire drought period. ChlF was more sensitive to drought monitoring than LCC. The original reflectance spectra and hyperspectral vegetation indices (SVIs) showed a strong correlation with LCC and ChlF. Reflectance in 540-560nm and 750-1100nm and selected SVI such as Simple Ratio (SR)752/690 can track drought responses effectively before leaves showed drought symptoms. Multivariate Linear Regression (MLR) and three machine learning algorithms, namely Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) were employed to develop models for estimating LCC and ChlF parameters. RF provided the best estimation accuracy for LCC compared to MLR, KNN and SVM, achieving an R<sup>2</sup> value of 0.895 for all LCC samples. The canopy layer significantly influenced the estimation accuracy of LCC, with the middle layer yielding the highest R<sup>2</sup> value. RF also demonstrated superior performance compared to MLR, KNN and SVM for estimating NPQ, Fv'/Fm', ETRmax, Fm', qL, qP, Y(II) and ETR, achieving R<sup>2</sup> value of 0.854 for NPQ, 0.610 for Fv'/Fm', 0.878 for ETRmax, 0.676 for Fm', 0.604 for qL, 0.731 for qP, 0.879 for Y(II), and 0.740 for ETR. Our results indicate that photosynthetic traits combined hyperspectral reflectance can monitor the effect of drought on trees effectively with significant potential for monitoring drought over large areas.
BibTeX
@inproceedings{li2025combine,
  author = {Linbao Li and Guiyun Huang and Jinhua Wu and Yunchao Yu and Guangxin Zhang and Yang Su and Xiongying Wang and Huiyuan Chen and Yeqing Wang and Di Wu},
  title = {{Combine photosynthetic characteristics and leaf hyperspectral reflectance for early detection of water stress}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1520304},
}
Figure